A relatively higher resolution digitized image (12) organized as a plurality of first kernel arrays (110), each with a plurality of pixels (37), is transformed into a corresponding relatively-smaller second kernel array (124) of a relatively-lower-resolution image (38) and an associated set (116, 118) of remaining pixels (37). Down-sampled pixels of the second kernel array (124) are generated from linear interpolation of original pixels (37) of the first kernel array (110). Associated interpolation coefficients incorporate perturbations to locations of the down-sampled pixels (37′) that are symmetric with respect to centers (128, 130) of the first (110) and second (124) kernel arrays. Down-sampled pixels (37′) of the second kernel array (124) can be recombined with the associated set (116, 118) of remaining pixels (37) to reconstruct the relatively higher resolution digitized image (12) substantially without loss of associated information, or used directly to reconstruct an approximation thereof, with associated encoding and decoding processes adapted to reduce the susceptibility of image reconstruction errors caused by subsequent image compression.